Zack Zeding Zhang

I'm a grad student in AI.
I focus on agent development, agent safety, and edge inference.

Looking for internships

me [at] zedingzhang.com

Open Source Contributions

On-device LLM inference engine

Fixed the QNN Embedding model export pipeline. Split the shared Embedding and standard LLM export flow by model type to resolve inconsistent model identification, I/O generation, and artifact path mapping; added Embedding-specific I/O and artifact mapping, and strengthened Qwen3 detection.
Added input-configuration validation for both Embedding and standard LLM export paths, then verified the change with Python syntax checks and a CMake build before it was merged upstream.

Secure sandbox for AI agents

Wrote Chinese and English OpenAI Agents SDK integration guides covering the SDK's built-in E2B client and a near-zero-change migration to E2B Cloud, with examples for Shell Agent, pause and resume, SWE-bench, and Code Interpreter.

AI security red-teaming platform

Fixed case conflicts in vulnerability rule directories. Removed 10 duplicate paths to prevent duplicate rule loading and detection issues on Windows and macOS; added case-folded detection, deterministic output, unit tests, and a CI gate to the Go yamlcheck tool, with 4,174 YAML files passing validation.

Python / DeepSeek V4 / Tkinter Desktop / CLI / Tree of Agents / Local Memory

Shipped a consent-first personal research and decision agent as both standalone CLI binaries and a native Tk desktop app. Its desktop workspace brings together files, tasks, relevant memory, agent activity, and approval-gated changes without starting a local web server.
Designed a bounded Tree of Agents mode that presents a preflight and runs only after explicit confirmation, coordinating up to two temporary workers before evidence validation.
Implemented approval-gated memory learning: feedback becomes a candidate card and changes future behavior only after review; relevance selects at most three local memories per task.

Docker / Git Worktree / Audit Logging / SHA-256 / CI

Built a Docker- and Git-worktree-based harness that isolates autonomous code modifications from source repositories, records integrity-verifiable audit artifacts, enforces fail-closed lifecycle states, and supports SHA-256-verified patch replay.
Verified the complete Docker bind-mount execution pipeline on Ubuntu CI with 26 integration and end-to-end tests passing.
Diagnosed and resolved Linux capability issues caused by cap-drop ALL, retaining only DAC_OVERRIDE for writable isolated worktrees.
Protected source repositories with detached worktrees and before/after fingerprints covering tracked files, untracked content, and symlink state.

WeChat Mini Program / JavaScript / glass-easel

Built a WeChat Mini Program for looking up character pronunciations in the Huazhou dialect (a Cantonese branch), supporting both Simplified and Traditional Chinese input with automatic script conversion.
Provides three pronunciation systems — Guangzhou Cantonese (Jyutping), Upper River, and Lower River Huazhou variants — with color-coded display for clarity.
Fully offline with zero cloud dependencies; dictionary covers 4,310 character entries with ~4,000 simplified-to-traditional mappings.